AGI/ASI Research 20250813: Computational Demonstration
Computational Demonstration
Using Python with PyTorch, I simulated a 30-layer sigmoid network (vanishing due to sig' <0.25) and a 40-layer linear chain with factor 1.2 (exploding).
Formula for Zero stabilization floors/caps the gradient products.
For Vanishing (early layers small):
Standard: [4.17e-20, 1.67e-19, ..., 0.22] (vanishes to near-zero).
FFZ: [0.001, 0.001, ..., 0.22] (floored at ε=0.001).
For Exploding (early layers large):
Standard: [1469.77, 1224.81, ..., 1.20] (explodes >1000).
FFZ: [1000.00, 1000.00, ..., 1.20] (capped at 1/ε=1000).
This shows FFZ preventing breakdowns, enabling training in deep nets.
Comments
Post a Comment